surogate/Surogate-3.5-9B

VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 15, 2026License:surogate-community-license-1.0Architecture:Transformer Featherless Exclusive Cold

Surogate-3.5-9B is a 9 billion parameter bilingual assistant developed by Invergent, optimized primarily for Romanian language instruction following, reasoning, and mathematical tasks. It features adaptive thinking capabilities, dynamically adjusting reasoning trace length based on problem complexity, while also supporting English. This model is designed for use on the Surogate platform and is compatible with the Qwen3.5 chat template.

Loading preview...

Surogate-3.5-9B: A Romanian-First Bilingual Assistant

Surogate-3.5-9B, developed by Invergent, is a 9 billion parameter bilingual assistant with a primary focus on the Romanian language. It is specifically optimized for Romanian instruction following, reasoning, and mathematical tasks, while maintaining strong support for English.

Key Capabilities & Features

  • Romanian-First Optimization: Excels in Romanian instruction following, reasoning, math, and orthographic correctness.
  • Adaptive Thinking Support: Utilizes explicit <think>...</think> reasoning traces in both Romanian and English. This feature dynamically adjusts the reasoning trace length based on the problem's complexity, offering brief deliberation for simple questions and more extensive reasoning for harder tasks.
  • Strong Performance: Achieves notable scores across various benchmarks, including:
    • Romanian Reasoning: 93.06 on RO thinking ARC and 75.61 on RO thinking MMLU.
    • Math: 75.42 on RO GSM8K (direct) and 91.89 on EN GSM8K (strict).
    • Instruction Following: 45.37 (prompt) / 68.72 (instruction strict) on RO IFEval.
    • Multilingual: Strong translation capabilities from EN to RO (57.27 chrF2) and RO to EN (64.67 chrF2).
  • Qwen3.5 Compatibility: Compatible with the Qwen3.5 chat template and tooling for ease of integration.

Use Cases & Considerations

This model is ideal for applications requiring high-quality Romanian language processing, particularly for tasks involving complex reasoning and mathematics. Developers should note that while English is supported, Romanian is the primary optimization target, and English performance may vary by task. The model was trained using Invergent's high-performance Surogate Trainer.

Finetunes
1 models